🧩 ModernBERT-base Fine-tuned for Harmful Prompt Classification

🧠 Model Overview

• Task: Harmful prompt detection (binary classification)
• Labels:
  • 1 → Harmful / Unsafe
  • 0 → Safe / Non-harmful

📊 Performance (Test Set)

⚙️ Training Details

• Dataset: allenai/wildguardmix (wildguardtrain subset)
• Split:
  • 80/20 train/test
  • 90/10 train/validation (from training set)
• Stratified on: prompt harm label, adversarial flag, and subcategory
• Optimizer: AdamW (8-bit)
• Learning Rate: 1e-4 (cosine schedule, 10% warmup)
• Batch Size: 96
• Max Sequence Length: 256 tokens
• Epochs: 3

🎯 Intended Use

• Harmful (1) — unsafe or toxic content
• Unharmful (0) — safe or benign content

⚠️ Disclaimer:
This model should not be deployed in production systems without additional evaluation and alignment with domain-specific safety and ethical guidelines.